Kernel-Based Methods for Vector-Valued Data with Correlated Components

Authors:
R. K. Beatson;W. zu Castell;S. J. Schrödl
Affiliations:
rick.beatson@canterbury.ac.nz;castell@helmholtz-muenchen.de and stefan.schroedl@gmx.net;-
Venue:
SIAM Journal on Scientific Computing
Year:
2011

Citing 6
Cited 0

A vector spline approximation

Journal of Approximation Theory
Generalized Hermite interpolation via matrix-valued conditionally positive definite functions

Mathematics of Computation
Radial Basis Functions

Radial Basis Functions
Handbook of Mathematical Functions, With Formulas, Graphs, and Mathematical Tables,

Handbook of Mathematical Functions, With Formulas, Graphs, and Mathematical Tables,
On Learning Vector-Valued Functions

Neural Computation
Universal Multi-Task Kernels

The Journal of Machine Learning Research

Quantified Score

Hi-index	0.00

Visualization

Abstract

This paper concerns kernel-based interpolation methods for vector data with correlated components. It gives conditions for a matrix kernel to be conditionally positive definite in an appropriate sense. The conditions allow construction of matrix kernels from nonsymmetric mixtures and scalings of scalar kernels. In particular the kernel used to model the influence of component $i$ on component $j$ can be different from that used to model the influence of component $j$ on component $i$. The vector modeling techniques considered are particularly appropriate when there are relatively few measurements of one quantity and relatively many of another “correlated” quantity. The paper concludes with some numerical tests on model problems.